You can automate most pre-delinquency reminders and early-stage collections conversations with an AI agent, as long as the contact rules run as code and not as prompt instructions. For debt collectors covered by US Regulation F, the safe harbour is seven or fewer calls about a debt in seven consecutive days, no calls for seven days after a conversation, and no contact before 8 a.m. or after 9 p.m. at the consumer's location. Add channel-level consent, a working opt-out and a fixed policy for what the agent may agree to, and it can remind, answer balance questions and set up payment arrangements. Everything else goes to a person.
Key takeaways
Regulation F, in force since 30 November 2021, presumes compliance at seven or fewer call attempts per person, per debt, in seven consecutive days, and a seven-day pause after any phone conversation about that debt.
Texts and emails do not count toward the call cap, but every electronic contact must carry a clear opt-out method, and a request to stop using a channel must be honoured.
An AI voice agent calling a mobile number needs prior express consent under the TCPA rules, and revocations must be honoured within ten business days at most: design for immediate.
Pre-due reminders are servicing, not collections. Keep them informational, stop them on payment, and hold the agent to arrangements your policy pre-approves.
The CFPB logged about 207,800 debt collection complaints in 2024; among electronic-communication complaints, 58% were about frequent or repeated messages. Frequency is where automation goes wrong first.
Which rules govern AI outbound collections in the US?
Four bodies of rules shape what an automated collections program may do: the Fair Debt Collection Practices Act (FDCPA), the CFPB's Regulation F that implements it, the FCC's TCPA rules on automated calls and texts, and state law. The CFPB describes Regulation F as the rule that prescribes Federal rules governing the activities of debt collectors, as that term is defined in the FDCPA. That definition matters. The FDCPA covers third-party collectors and debt buyers; a creditor collecting its own debt under its own name is generally outside it, but still answers to unfair, deceptive and abusive practice rules, state statutes and its prudential regulator. Treating the Regulation F standards as the operating floor is the conservative choice either way.
The TCPA layer applies to everyone. Under 47 CFR 64.1200, no one may place a call using an automatic dialling system or an artificial or prerecorded voice to a mobile number without the called party's prior express consent, other than for emergencies. Treat an AI-generated voice as an artificial voice for this purpose. The same rule says revocation requests made in any reasonable manner must be honoured within a reasonable time not exceeding ten business days. Honour them the same minute.
The enforcement picture shows where programs fail. The CFPB's 2025 FDCPA annual report counts approximately 207,800 debt collection complaints in 2024, seven percent of all complaints it received. Among complaints about electronic communications, 58% cited frequent or repeated messages and 32% said the collector kept trying after being told to stop. Those are the two failure modes an automated cadence produces at scale if nobody caps it.
How often and when may an AI agent contact a customer?
The answer for phone calls is the 7-in-7 rule. Regulation F section 1006.14 presumes a collector complies if it places calls about a particular debt neither more than seven times within seven consecutive days, nor within a period of seven consecutive days after having had a telephone conversation with the person about that debt. The CFPB's own FAQs add two details that trip up automation teams: the presumption applies per person, per debt, regardless of how many telephone numbers are associated with a particular person, and the call-frequency prohibition does not apply to other media such as text messages, email or social media. So a customer with three phone numbers still has one counter per debt, and your SMS cadence needs its own limits.
Timing is set by section 1006.6. Absent knowledge to the contrary, a time before 8:00 a.m. and after 9:00 p.m. local time at the consumer's location is inconvenient. Your scheduler has to resolve each account to a time zone from address and area code, and hold sends when the two disagree. The same section requires every email or text to include a clear and conspicuous statement describing a reasonable and simple method to opt out of further electronic communications to that address or number. Section 1006.14(h) then says a collector must not communicate through a medium the person has asked it not to use. A customer who texts STOP has opted out of texts; whether a voice call an hour later is acceptable is a decision your own rules must make.
Two more clocks run in the background. The FDCPA at 15 U.S.C. 1692c says that if a consumer notifies a collector in writing that they refuse to pay or want communication to stop, the collector shall not communicate further about that debt, with narrow exceptions. Section 1006.34 requires validation information in the initial communication or within five days of that initial communication, opening a 30-day validation period during which the consumer can dispute the debt. Section 1006.18(e) also requires the initial communication to disclose that the collector is attempting to collect a debt.
The design consequence is simple. Frequency counters, time-zone checks, consent flags, opt-out state, cease flags and disclosure insertion are all deterministic. They belong in a pre-send check that runs before every message on every channel. The language model never gets a vote on whether a contact is allowed; its job starts after the check passes.
What cadence works for pre-delinquency reminders versus early-stage collections?
Pre-due reminders and early-stage collections are different programs and should be built as such. A pre-due reminder is a servicing message to a customer in good standing: amount, date, a one-tap way to pay, no consequence language. Early-stage collections starts the day a payment fails, and it is a conversation with a cadence behind it, where the goal is to learn why the payment failed and fix it inside policy.
The economics argue for investing in the early window. The New York Fed's Q2 2026 Household Debt and Credit Report put 4.7% of outstanding debt in some stage of delinquency, with new delinquencies for auto loans and credit cards still at elevated levels. And the Federal Reserve's 2025 household well-being survey found that only 63 percent of adults would cover a $400 emergency expense using cash or its equivalent. Roughly a third of your book cannot absorb a surprise, so a hardship path has to exist inside the early-stage flow.
Stage | Goal | Channels | Suggested cadence | Hard stops |
|---|---|---|---|---|
Pre-due (T-5 to T-1) | Prevent the miss | Email, SMS if consented | One email at T-5, one SMS at T-1 with pay link | Stop on payment or autopay enrolment |
Day 1 to 7 past due | Fix the failed payment | Email, SMS, in-app | Failure notice day 1, conversational SMS day 3, email with options day 7 | Halt cadence on any reply; suppress during an active conversation |
Day 8 to 30 past due | Reach the customer and agree a plan | Adds voice if consented | At most one call attempt per day, three per week, inside 8 a.m. to 9 p.m. local | 7-in-7 counter per debt; seven-day pause after any conversation |
Day 31 to 60 past due | Arrangement or hand-off | Voice, SMS, email | Agent leads first contact, human owns follow-up when hardship or dispute appears | Cease and attorney flags block all outbound |
Day 60 and later | Human-led recovery | Human decides | AI provides context and takes inbound only | No automated outbound without policy sign-off |
Two guardrails make any cadence safe. Suppress-on-active-conversation: if the customer is mid-thread with the agent or a human, the scheduled day-7 message does not fire. Suppress-on-resolution: once a payment posts or an arrangement is confirmed, every pending contact for that account cancels on every channel. Both are state checks, not judgement calls. Keep the counts well under the legal ceiling: a customer who hears from you seven times in a week is a complaint in progress.
What may an AI collections agent do, and what must it never do?
An AI agent may do the work a well-supervised first-line collector does: state the balance and due date from the system of record, explain how the amount was calculated, take a payment, set up an arrangement from a pre-approved menu, record a promise to pay, capture a dispute and hand it off, and process an opt-out on the spot. Each action maps to a tool with scoped permissions, so the agent can only read the account it is talking to and only write outcomes policy allows.
What it must never do follows directly from section 1006.18. A collector must not use any false, deceptive or misleading representation in collecting a debt, and must not threaten to take any action that cannot legally be taken or that is not intended to be taken. That gives an AI agent four hard rules. It never invents a settlement, discount or deferral that is not on its menu. It never implies credit reporting, legal action or fees unless your policy confirms that outcome is real and scheduled. It never overstates the amount, the legal status of the debt, or who it is. And it never keeps collecting after a dispute, a cease request or notice of attorney representation. The first three are prompt and tool-permission problems. The last is a state problem for the pre-send check.
Hardship decides whether the program helps or harms. When a customer says they lost a job or had a medical bill, the agent should acknowledge it, stop the cadence, and offer only the options that customer is eligible for. Define eligibility in advance: which balances qualify for a two-instalment split, how long a deferral may run, when a partial settlement is even on the table. Anything outside that box is a warm hand-off with the transcript, not a creative counter-offer. Record the hardship state so a future contact does not restart the clock.
Escalation triggers should be explicit and boring: a dispute of the debt or amount, a validation request, a cease or a lawyer's name, a vulnerability signal, a threatened complaint, or a hostile turn. An agent that hands off cleanly on the hard ten percent earns the trust to run the routine ninety.
What this looks like in practice: a worked example
Take a US consumer lender with a direct-debit book that wants to automate pre-due reminders and days 1 to 30 of collections on outbound across SMS, email and voice. Here is how the pieces fit on Lorikeet, and where the platform's job ends.
The contact rules go into deterministic structured workflows: per-debt call counters, the seven-day post-conversation pause, time-zone resolution against the customer's location, channel consent flags, and the cease and attorney blocks. Those run before any message is generated. The conversation runs as a natural-language workflow with a fixed arrangement menu and scoped tools to read the balance, take a payment and write a promise-to-pay date back to the servicing system. Outbound guardrails check each generated message for disclosure presence, threat language and off-menu offers before it sends. Lorikeet's outbound is consent-first: the agent identifies itself as an AI, states the required collection disclosure, and works only against numbers and addresses your consent records allow.
Before go-live, the compliance team runs simulations against the bad paths: an eighth call attempt, a send at 7:40 a.m. local, a STOP followed by a scheduled voice call, a hardship message answered with a template. Sign-off is on observed behaviour. After go-live, Coach reviews 100% of conversations against the same criteria, replacing sampled QA for the disclosure, timing and tone checks. Controls and certifications are on the trust page and quality assurance page.
The published proof for the outbound mechanics comes from Carmoola, an FCA-regulated UK car finance lender, whose Lorikeet outbound agent resolved 90% of its outbound conversations end to end and lifted a conversion metric by 60%. That deployment was proactive re-engagement under UK rules, not US collections, so read it as evidence that the outbound channel resolves conversations at scale, not as a collections benchmark. Pricing is per resolution, about $0.80 for chat, email or SMS and about $1.20 for voice, and unresolved tickets cost nothing. See the pricing page, and the financial services page for how the same setup extends to payment-failure support.
The limitation is worth stating plainly. Lorikeet is designed to support compliance programs; it does not make you compliant. You write the contact policy, the arrangement menu and the eligibility rules; your counsel decides how the FDCPA and state rules apply. Consent capture happens upstream in your onboarding systems, and the agent can only honour what those records say. If you want to test your own contact rules against the bad-path scenarios above, book a session and bring the policy document.
What still needs a human
Several parts of collections should stay with people, and a good deployment routes them there automatically: disputes and validation requests, because the response has legal content and a 30-day clock; cease notices and attorney representation, because permitted responses narrow to almost nothing; hardship outside the arrangement menu and any sign of vulnerability, because discretion is the job; settlement beyond pre-approved parameters; complaints; and accounts past 60 days, where the questions are about legal strategy rather than reminders.
People also own the program. Someone has to write and version the contact policy, approve the arrangement menu, read the simulation reports, sample Coach's QA findings to check the reviewer, and update the rules when the CFPB, the FCC or a state legislature moves. The agent executes policy; it does not author it. For the broader pattern of splitting deterministic controls from language-model judgement in regulated support, see AI customer support in fintech and involuntary churn prevention.
Build the rules layer first, keep the agent's authority to a written menu, and prove the bad paths fail safely before the first live send. Do that and an AI agent will run reminders and early-stage work more consistently than a tired human on a Friday evening, with a record of every contact your regulator can read.
